Detecting human-to-human transmission of avian A (H5N1) influenza

Page 1

Page 1 of 14

Suggested citation for this article: Yang Y, Halloran ME, Sugimoto J, Longini, Jr IM. Detecting

human-to-human transmission of avian influenza A (H5N1). Emerg Infect Dis. 2007 Sep; [Epub

ahead of print]

Detecting Human-to-Human Transmission

of Avian Influenza A (H5N1)

Yang Yang,* M. Elizabeth Halloran,*† Jonathan Sugimoto,*† and Ira M. Longini, Jr.*†

*Fred Hutchinson Cancer Research Center, Seattle, Washington, USA; and †University of Washington, Seattle,

Washington, USA

Highly pathogenic avian influenza A (HPAI) subtype H5N1 has caused family case clusters, mostly in

Southeast Asia, that could be due to human-to-human transmission. Should this virus, or another

zoonotic influenza virus, gain the ability of sustained human-to-human transmission, an influenza

pandemic could result. We used statistical methods to test whether observed clusters of HPAI (H5N1)

illnesses in families in northern Sumatra, Indonesia, and eastern Turkey were due to human-to-human

transmission. Given that human-to-human transmission occurs, we estimate the infection secondary

attack rates (SARs) and the local basic reproductive number, R

. We find statistical evidence of human-

to-human transmission (p = 0.009) in Sumatra but not in Turkey (p = 0.114). For Sumatra, the estimated

household SAR was 29% (95% confidence interval [CI] 15–51). The estimated lower limit on the local R

was 1.14 (95% CI 0.61–2.14). Effective HPAI (H5N1) surveillance, containment response, and field

evaluation are essential to monitor and contain potential pandemic strains.

Highly pathogenic avian influenza A (HPAI) subtype H5N1 is repeatedly crossing the

species barrier to humans. Since December 2003, a total of 291 cases of HPAI (H5N1) have been

reported in humans, resulting in 172 deaths (i.e., 59% case-fatality ratio) in 12 countries, mostly

in Southeast Asia (1). Among these cases, 31 family clusters have been documented, ranging in

size from 2 to 8 family members. How many of these clusters are due to a common avian source

and how many are due to human-to-human transmission are important facts to determine. Should

one of these HPAI (H5N1) strains gain the capacity for sustained human-to-human transmission,

the resulting outbreak, if not contained, would spread worldwide through the global

transportation network more rapidly than adequate supplies of vaccine matched to the new

Page 2

Page 2 of 14

variant could be manufactured and distributed (2,3). We analyzed data from 2 of the largest of

the familial clusters to ascertain if human-to-human transmission took place, and if so, how

transmissible the strain was.

Methods

May 2006 Human Avian Influenza Family Cluster, Indonesia

During late April and early May 2006, a cluster of 8 cases of HPAI (H5N1) was detected

and investigated by the Indonesian public health surveillance system in northern Sumatra (4–6).

All case-patients were members of the same extended family. Seven of them resided within 3

adjacent houses in the village of Kubu Sembilang. The remaining patient resided with his

immediate family in the village of Kabanjahe (≈10 km away).

The index patient was a 37-year-old woman, thought to have been exposed to dead

poultry and chicken fecal material before onset of illness. She also reportedly maintained a

market stall that sold live chickens. Although her illness was not confirmed to have been caused

by avian influenza (H5N1), her death on May 5, 2006, is suspected to be the result of HPAI

(H5N1) infection because of her reported symptoms, illness progression, and prior contact with

diseased or dead poultry.

Twenty members of her extended family are suspected to have been in contact with her,

many during a family gathering on April 29, 2006 (7). At that time, she was manifesting

symptoms (i.e., she had a heavy cough, was severely ill, and was prostrate). That night, 9 of

these members slept in the same small room as she did (indicated by a black triangle in

Appendix Figure 1). Of these 9 family members, 2 of her sons (15 and 17 years of age) and her

25-year-old brother, who lived in Kabanjhe, became ill in the next 3 weeks. The sons died. The

brother was the only person from this family cluster to recover.

Of the remaining 11 family members, 4 became ill and died. The 29-year-old sister of the

index patient, who lived in an adjacent house, became ill after she provided direct personal care

to her ill sister (7). The 18-month-old daughter of this sister also became ill after she was in the

presence of the index patient with her mother. The 10-year-old nephew of the index patient, who

lived in the other house adjacent to hers, became ill after he attended the family gathering and

frequently visited his aunt’s house. The nephew’s father became ill after he personally cared for

Page 3

Page 3 of 14

his son. The possibility that HPAI (H5N1) was transmitted from the nephew to his father is also

supported by genetic sequencing data (4). Though symptoms did not develop in the mother of the

nephew, she was directly exposed to her husband during his illness. All case-patients, except for

the index patient, were confirmed as influenza (H5N1) positive by PCR. The nephew’s mother

was confirmed as influenza (H5N1) negative. As an intervention, 54 surviving relatives and

close contacts were identified and placed under voluntary quarantine (7). All of these persons,

except for pregnant women and infants, received oseltamivir prophylactically.

December 2005 Human Avian Influenza Family Cluster, Eastern Turkey

From December 18, 2005, (8) to January 15, 2006 (9), a cluster of 8 confirmed influenza

(H5N1) cases was detected in Dogubayazit District in eastern Turkey (Appendix Figure 2) (10–

13). These case-patients were among 21 members of 3 households located within 1.5 km of each

other (14). All confirmed case-patients were hospitalized after onset of symptoms (9). Four of

the confirmed case-patients died; the other 4 recovered (9). Ten of the remaining 14 household

residents were hospitalized with avian influenza-like symptoms but were never confirmed to be

infected with influenza (H5N1) (9). All but one of the hospitalized residents were children (6–15

years of age) (9).

Before onset of symptoms, 4 children from 1 household, 3 of whom had confirmed cases

(including the index patient), were reported to have had close contact with the dead bodies of

sick chickens (15). The 2 confirmed case-patients in the second household reportedly slaughtered

a duck together on January 1, 2006, at the beginning of a die-off in the household’s flock (14).

Two of the remaining confirmed case-patients lived in the third household and had no history of

contact with sick or dying poultry. The remaining confirmed case occurred in a fourth residence

located near the first household (10), but because we lacked information on the number of

household members and the case-patient’s exposure history, we excluded it from these analyses.

Most, if not all, of the 21 residents attended a dinner hosted by the family of the index patient on

December 24, 2006, while he was symptomatic (8).

Statistical Methods

We used a previously developed statistical transmission model (16,17) to test whether

human-to-human transmission occurred, and if it did, to estimate transmission parameters. In the

model, persons mix with one another in households and between households. In addition, we

Page 4

Page 4 of 14

include a common source of infection due to zoonotic exposure. Mathematical and statistical

details are given in the online Technical Appendix (available from

www.cdc.gov/EID/content/13/7/07-0111-Techapp.pdf).

Model of Probability of Transmission

We define p

as the probability that an infectious household member infects another

household member in 1 day. If the distribution of the infectious period is known, we can obtain

the household secondary attack rate (SAR

) from p

, defined as the probability that an infectious

household member infects another household member over his or her infectious period.

Similarly, we define the daily transmission probability (p

) and the community SAR (SAR

) for

between household spread. Finally, we define the daily probability (b) that any person is infected

from a zoonotic source. The contact structure used for parameter estimation is shown in the

Figure. We assume that the distributions of the incubation and infectious periods are

predetermined by the investigator.

We establish the likelihood function for each person and then for the whole population

for statistical inference. The likelihood function for a person is equivalent to the probability of

observing the realized data on that person throughout the outbreak. The likelihood function for a

person labeled i is built with the following steps: 1) Obtain the probability that person i is

infected by an infectious source labeled j on day t, given person i is not infected up to day t – 1.

If source j is a person, this probability is p

, for the same household, or p

for exposure in the

community, multiplied by the probability of person j being infectious on day t. The probability of

person j being infectious on day t is derived from the symptom-onset day of person j and the

distribution of the infectious period. If source j is zoonotic, the infection probability is b. The

probability of escaping infection is simply 1 minus the corresponding probability of infection. 2)

Take the product of the probabilities obtained in step 1 over all humans and zoonotic sources j to

obtain the probability of person i escaping infection by any infectious source on day t. 3) Take

the product of the probabilities obtained in step 2 over all days before and including day t to

obtain the probability of person i escaping infection up to day t. 4) If person i is not infected by

the end of the outbreak, the likelihood function for person i is the product of the probabilities of

person i escaping infection up to the last day of observation. 5) If person i is observed to have

symptom onset on day t and the infection time is known to be t, the probability of the data

Page 5

Page 5 of 14

regarding person i is the product of 3 pieces of information: a) the probability of person i

escaping infection up to day t – 1, b) the probability that person i is infected on day t, and c) the

probability that the duration of the incubation period is t – t. Because we do not observe the

infection time, the likelihood function for person i is obtained by summing the above product, a –

c, over all potential values of t.

The likelihood function for the whole population is the product of all the individual

likelihood functions. In the event that human-to-human transmission occurs, SAR estimates are

used to estimate the local basic reproductive number (R

), which is defined as the average

number of secondary cases infected by a typical index case-patient in the beginning of the

outbreak (online Technical Appendix). There is potential for sustained transmission if R

is >1. If

human-to-human transmission is determined to be occurring, then the above parameters are

estimated from the symptom dates and contact information from the population under study.

Data on exposed persons who do not become ill form an important component of the inference

procedure.

Statistical Test

We set up a statistical test with the null hypothesis being that no human-to-human

transmission occurs, that is, p

= p

= 0. The alternative hypothesis is either p

is not equal

to 0, or both are not equal to zero. The test statistic we use is proportional to the ratio of the

maximum value of the likelihood function assuming the null hypothesis is true (null likelihood)

and the maximum value of the likelihood function at the estimated parameter values (full

likelihood).

Specifically, we define the likelihood ratio test statistic as –2 log (the null likelihood

function divided by the full likelihood function). If no human-to-human transmission occurs, the

2 likelihood functions would be roughly equal, and we expect to see a likelihood ratio close to 1,

and, thus, a likelihood ratio statistic close to 0. A large value of the likelihood ratio statistic is

evidence of deviation from the null hypothesis. The question is how to obtain a reference set of

the likelihood ratio statistic values that we would see under the null hypothesis. Given no human-

to-human transmission, all the observed case-patients must have been infected by the zoonotic

source. Since the exposure to the zoonotic source is assumed constant for each person on each

day, the null likelihood function will not change if we reassign the infection and symptom status

Page 6

Page 6 of 14

of the observed case-patients to a different group of people in the population. By performing

such reassignment many times, we obtained a collection of datasets that were each equally likely

to have been observed had there been no human-to-human transmission. The values of the

likelihood ratio statistic calculated from these datasets form the null distribution for statistical

testing. This method is referred to as a permutation test. The p value is given by the proportion of

the reference values that are equal to or larger than the observed likelihood ratio statistic value.

More technical details are given in the online Technical Appendix.

The probability of infection by the zoonotic source may not be estimable together with

SAR

or SAR

from an observed cluster. In such a situation, a statistical test of the occurrence of

human-to-human transmission is still meaningful because the likelihood ratio test statistic is still

estimable from the permuted datasets.

Data Required

A list of the inputs that are required for estimation and statistical testing are listed in the

Table. Three categories of input parameters are required for this estimation model: outbreak-

wide, individual level, and analysis parameters. The duration of the outbreak, the duration of the

incubation period for the pathogen, and the minimum and maximum durations of the infectious

period for the pathogen are the required outbreak-wide inputs. For each person, their residential

location (neighborhood and household), their demographic characteristics (sex and age), and

whether they were a case-patient or not are required input parameters. Case-patients require

additional input of their illness-onset dates, types of outcome, outcome dates, and whether or not

they are the index patient in the outbreak. Hospitalization and treatment dates (considered

prophylactic for nonpatients) are optional input parameters for each person. For each person who

visits another residence during the outbreak period, his or her identifiers, the neighborhood and

household visited, and the start and end dates of the visit are required inputs. Analysis-related

inputs include the last date of community exposure to potential common sources of infection, the

last date of observation, and inputs for R

estimation (mean number of residents per household

and mean number of out-of-residence contacts per person per day). An expanded version of the

model will require the input of other exposure information such as from schools or hospitals.

Page 7

Page 7 of 14

Results

For the outbreak in Indonesia, Online Figure 1 shows that the incubation period had a

probable range of 3–7 days and the infectious period, a probable range of 5–13 days. Thus, we

let the incubation period have a uniform distribution of 3–7 days (mean 5 days) and the

infectious period a uniform distribution of 5–13 days (mean 9 days). For the data shown in

Online Figure 1, only the household SAR (SAR

) can be estimated. We determine that human-

to-human spread did occur by rejecting the null hypothesis of no human-to-human transmission

(p = 0.009). The estimated household SAR is 0.29 (95% confidence interval [CI] 0.15–0.51).

Thus, a single infected person in a household infected another household member with the

probability of 0.29. The average household size for rural Indonesia is ≈5 people. Because we do

not have an estimate of the community SAR, we have an estimate of the lower limit of the local

i.e., 1.14 with a 95% CI of 0.61–2.14. A sensitivity analysis on the distribution of the

incubation and infectious period shows that the test and estimates for SAR

and R

are

insensitive to uncertainty about these distributions within plausible ranges.

For the outbreak in Turkey, all the parameters are estimable, but we do not reject the null

hypothesis of no human-to-human transmission (p = 0.114). Our estimate of the daily probability

of infection from the common source is 0.011 (95% CI 0.005–0.025).

Discussion

We have presented statistical evidence that the strain of HPAI (H5N1) that caused the

family cluster of human cases in northern Sumatra was spread from human to human and that the

household SAR was 29%. This household SAR is similar to statistical estimates for

interpandemic influenza A in the United States (12.7%–30.6%) (18,19). The mean incubation

period of this strain appears to have been ≈5 days, nearly twice as long as for past pandemic

strains and current interpandemic strains of influenza. The CI for the estimated lower bound for

the local R

covers 1. Therefore, even though we determined that human-to-human transmission

probably occurred, whether the virus was capable of sustained human-to-human transmission is

not clear. This virus may have required very close human contact to be transmitted. Even with no

intervention, the finding that R

= 1.14 indicates that the chance that a single introduction would

result in any further spread is ≈12%. In addition, the reported prophylactic use of oseltamivir

Page 8

Page 8 of 14

may have played some role in limiting further spread. We did not find statistical evidence of

human-to-human spread for the outbreak in eastern Turkey. This does not mean that no low-level

human-to-human spread occurred in this outbreak, only that we lack statistical evidence of such

spread. The power would be too low to detect such spread for an outbreak with 7 total cases and

small SARs (17).

We did not consider the role of heterogeneity—such as age, sex, treatment status, or

quarantine—in transmission. The parameters could be made to be functions of time-dependent

covariates, as we have done with similar models (16,19,20). We can easily extend the model

used here for covariates; however, we must have sufficient data to support such models.

Computer simulations have shown that the targeted use of influenza antiviral agents

could be effective in containing a potential pandemic strain of influenza at the source (21,22), if

initiated within 3 weeks of the initial case in the community, and if the R

is <1.8. This strategy,

known as targeted antiviral prophylaxis, involves treating identified index patients in a mixing

group and offering a single course of prophylaxis to the contacts of these index patients in

predefined close contact groups, i.e., households at a minimum but also possibly neighborhood

clusters, preschool groups, schools, and workplaces. In addition, the voluntary household

quarantine of suspected close contacts of case-patients was recommended. Targeted antiviral

prophylaxis at the household and neighborhood cluster level was carried out for the outbreak in

Sumatra.

Ascertaining whether a potential pandemic strain of influenza is capable of sustained

human-to-human transmission and estimating key transmission parameters are important. To

estimate more than the household SAR, more detailed community data need to be collected. This

would include a complete census of potentially exposed households and persons in the area

where immediate transmission could occur from both potential zoonotic and human sources.

Such data would enable estimation of important parameters and a more complete estimate of the

rather than just the lower limit.

We have developed a software application, TRANSTAT, for implementing these

analyses. This application provides a stand-alone environment for the entry, storage, and analysis

of data from outbreaks of acute infectious diseases. A partial list of the input information is given

in the Table. The statistical methods presented here can be applied to the data along with several

Page 9

Page 9 of 14

standard epidemiologic tools. This information system would allow for real-time analysis and

evaluation of control measures for an outbreak. We would encourage outbreak investigators to

use this tool, taking care to input data on the exposed nonpatients as well as case-patients. The

authors will provide a link to this software upon request.

This work was supported by the National Institute of General Medical Sciences MIDAS grant U01-

GM070749 and National Institute of Allergy and Infectious Diseases grant R01-AI32042.

Dr Yang is a staff scientist in the Biostatistics and Biomathematics Program in the Division of Public

Health Sciences at the Fred Hutchinson Cancer Research Center, Seattle, Washington. His primary research interest

is in the statistical and mathematical analysis of infectious disease data and intervention studies.

References

1. World Health Organization. Confirmed human cases of avian influenza A (H5N1). Epidemic and

pandemic alert and response. 2007 Apr 11. [cited 2007 Apr 25]. Available from

http://www.who.int/csr/disease/avian_influenza/country/cases_table_2007_04_11/en/index.html

2. Webby RJ, Webster RG. Are we ready for pandemic influenza? Science. 2003;302:1519–22. Medline

3. Stohr K. Avian influenza and pandemics—research needs and opportunities. N Engl J Med.

2005;352:405–7. Medline

4. Butler D. Family tragedy spotlights flu mutations. Nature. 2006;442:114–5. Medline

5. World Health Organization. Avian influenza—situation in Indonesia—update 12. Jakarta, Indonesia.

Report no. 12. 2006 May 19. [cited 2006 Sep 22]. Available from

http://www.who.int/csr/don/2006_05_16b/en/index.html

6. Soebandrio A. Indonesia: avian flu from Indonesia’s experience and perspectives. Pandemic

Preparedness and Infection Control. 2006 July 13–14.

7. World Health Organization. Avian influenza—situation in Indonesia—update 16. Jakarta, Indonesia.

Report no. 16. 2006 Jun 2. [cited 2006 Sep 16]. Available from

http://www.who.int/csr/don/2006_05_31/en/index.html

8. Recombinomics. H5N1 bird flu pandemic phase evolution. 2006 Jun 5. [cited 2007 Mar 16]. Available

from http://www.recombinomics.com/news/06050601/h5n1_phase_evolution.html

9. Recombinomics. Timeline for Kocyigit Ozcan family clusters in Dogubeyazit. January 22, 2006. [cited

2006 Mar 16]. Available from

http://www.recombinomics.com/news/01220601/h5n1_kocyigit_ozcan_timeline.html

Page 10

Page 10 of 14

10. World Health Organization. Avian influenza, situation in Turkey, update 2. Ankara, Turkey. Report

no. 2. Epidemic and pandemic alert and response. 2006 Jan 9. [cited 2007 Feb 27]. Available

from http://www.who.int/csr/don/2006_01_09/en/index.html

11. World Health Organization. Avian influenza, situation in Turkey, update 4: sequencing of human

virus. Ankara, Turkey. Report no. 4. Epidemic and pandemic alert and response. 2006 Jan 12.

[cited 2007 Feb 27]. Available from http://www.who.int/csr/don/2006_01_12/en/index.html

12. World Health Organization. Avian influenza, situation in Turkey, update 5. Ankara, Turkey. Report

no. 5. Epidemic and pandemic alert and response. 2006 Jan 16. [cited 2007 Feb 27]. Available

from http://www.who.int/csr/don/2006_01_16/en/index.html

13. Lauer C. Turkey’s bird flu outbreak: one year later. Turkish Daily News. 2007 Jan 12. [cited 2007

Mar 20]. Available from http://www.turkishdailynews.com.tr/article.php?enewsid=63680

14. World Health Organization. Shindo N. Avian influenza outbreak response in Turkey, 2006. Epidemic

and pandemic alert and response. [cited 2007 Mar 20]. Available from http://www.col.ops-

oms.org/servicios/influenza/reunion/docs/trad/10_turkey_summary.pdf

15. World Health Organization. Avian influenza, situation in Turkey. Ankara, Turkey. Epidemic and

pandemic alert and response. Report no. 1. 2006 Jan 5. [cited 2007 Feb 27]. Available from

http://www.who.int/csr/don/2006_01_05/en/index.html

16. Yang Y, Longini IM, Halloran ME. Design and evaluation of prophylactic interventions using

infectious disease incidence data from close contact groups. Applied Statistics. 2006;55:317–30.

17. Yang Y, Longini IM, Halloran ME. A resampling-based test to detect person-to-person transmission

of infectious disease. Annals of Applied Statistics. 2007;1:211–28. Available from

http://projecteuclid.org/dpubs/repository/1.0/disseminate?handle=euclid.aoas/1183143736&view

=body&content-type=pdfview_1

18. Longini IM, Koopman JS, Monto AS, Fox JP. Estimating household and community transmission

parameters for influenza. Am J Epidemiol. 1982;115:736–51. Medline

19. Longini IM, Koopman JS, Haber MJ, Cotsonis GA. Statistical inference for infectious diseases: risk-

specific household and community transmission parameters. Am J Epidemiol. 1988;128:845–59.

Medline

20. Rampey AH, Longini IM, Haber MJ, Monto AS. A discrete-time model for the statistical analysis of

infectious disease incidence data. Biometrics. 1992;48:117–28. Medline

Page 11

Page 11 of 14

21. Longini IM, Nizam A, Xu S, Ungchusak K, Hanshaoworakul W, Cummings DA, et al. Containing

pandemic influenza at the source. Science. 2005;309:1083–7. Medline

22. Ferguson NM, Cummings DAT, Cauchemez S, Fraser C, Riley S, Meeyai A, et al. Strategies for

containing an emerging influenza pandemic in Southeast Asia. Nature. 2005;437:209–14.

Medline

Address for correspondence: Ira M. Longini, Jr, Program in Biostatistics and Biomathematics, Fred Hutchinson

Cancer Research Center, Seattle, WA 98109-1024, USA; email: longini@scharp.org

Table. Parameters and data used in analysis